The right chatbot for a business is chosen on three criteria, in this order: what customers actually ask, which sources the assistant can answer from, and which actions it has to take on its own (booking, taking a request, checking a status). The technology — rules, generative AI or a combination — comes only after you've answered these questions. The guide below walks through the decision exactly the way we do it in Superior Solutions projects.
The three types of chatbot and when each makes sense
Rule-based chatbot (menus and fixed flows). The user chooses from options and the answers are written in advance. It makes sense when there are few, predictable questions: opening hours, address, “how do I book”. Advantage: it never goes wrong outside the script. Limitation: any question phrased differently hits “I didn't understand”.
AI assistant that answers from company information. For each question, the language model receives the relevant passages from approved sources (website, documents, policies, product sheets) and phrases the answer from them — not from the model's general memory. It makes sense when questions vary but the answers already exist somewhere in the company. Limitation: answer quality depends directly on the quality and freshness of the sources.
Assistant with actions (hybrid). Besides answering, it can check a calendar, create a request in the CRM, look up an order status or hand the conversation to a colleague with the full context. It makes sense when the goal is not only to inform but to complete a step. This is also where most project decisions appear: what it can do alone, what requires identifying the person and what must be approved by a human.
Step 1: inventory the real questions
Before any demo, collect the conversations from the last few weeks: email, phone (the team's notes), WhatsApp, the contact form, social media comments. Group them by topic and note three things for each topic:
- How often it comes up — a frequent topic justifies the effort of solving it automatically; a rare one doesn't.
- Whether the answer is public or personal — “what warranty do you offer” is public; “where is my order” requires identification.
- Whether the answer ends in an action — pure information, a booking, a quote request, a complaint.
The result is a short list of what the assistant must do in the first version. Anything not on the list can be handled with a polite handover to the team. A chatbot that tries to cover everything from the start answers everything badly.
Step 2: decide the sources it answers from
An AI assistant is exactly as good as its sources. For each topic on the list, decide where the answer comes from:
- Public content (service pages, FAQs, delivery and returns policies): can be used directly, but must be up to date. An old page produces old answers.
- Internal documents (procedures, technical sheets, framework contracts): must be cleaned of information you don't want customers to see and marked with the date of the last revision.
- Systems with personal data (CRM, orders, bookings): access happens only after identifying the person, with rights limited to their own data. This is where we also involve the data protection officer.
A principle we follow in every implementation: the assistant must be able to say “I don't know” and hand over. A confidently invented answer costs more than a handover to a colleague.
Step 3: define the actions and the limits
For assistants with actions, write three explicit lists:
- What it does alone, without confirmation — for example, sending the booking link or confirming opening hours.
- What it does only after the user confirms — creating the booking, recording the quote request.
- What it never does — it doesn't change prices, cancel paid orders or promise deadlines it can't verify.
The handover to the team is also an action and must be designed just as carefully: the colleague receives a summary of the conversation, the data already collected and the reason for the handover — not just “a customer wants to talk to someone”.
The channel: website, WhatsApp, Messenger or internal
Website. The simplest to start with, no external approvals. Good for information and taking requests.
WhatsApp. Heavily used in Romania, especially in hospitality, services and retail. It is implemented through the WhatsApp Business Platform, and the costs follow their own logic: since 1 July 2025, Meta charges per delivered template message (marketing, utility, authentication), while regular messages sent within the 24-hour window opened by the customer are free. Conversations started from a click-to-WhatsApp button or ad get a free 72-hour window. In practice: an assistant that replies to customers who write first costs little; one that sends proactive notifications must be budgeted by volume.
Messenger / Instagram. Useful if your audience already contacts you there; the same principles as WhatsApp, under Meta's platform rules.
Internal (for the team). The assistant answers colleagues from procedures and documents. It is often the quickest win, because it doesn't involve customers and the risks are lower.
What to check before launch
- The test question set: 30–50 real questions, phrased the way customers write them (with typos, abbreviations, missing diacritics), with the expected answer for each. We run it on every change to the sources.
- Behaviour on unknowns: what it says when it can't find the answer and how it hands over.
- Identification: how it verifies the person before showing personal data.
- The conversation log: where it is stored, who can see it, how long it is kept and how you use it to improve the sources.
- Informing the user: that they are talking to an automated assistant and how they can reach a human.
Mistakes we see often
- Launching without a list of real questions — the assistant is judged on impressions, not cases.
- Neglected sources: contradictory pages, old prices, unrevised procedures.
- No handover with context — the customer repeats everything they already wrote.
- A single channel chosen by fashion, not by where customers actually write.
- Nobody responsible for reviewing conversations after launch.
What drives duration and cost
There is no standard price, because there is no standard chatbot. The difference comes from the number of sources and how clean they are, integrations with internal systems (calendar, CRM, orders), the channels, the need to identify the person and the conversation volume (which drives the cost of the AI model and, on WhatsApp, of template messages). A first version limited to information and taking requests, on a single channel, is usually the shortest path to a result you can evaluate.
See what a project includes on the Chatbots & virtual assistants page or, if you want to go beyond conversation, AI agents.